Model Card
The document that spells out what a model can and can't do
- A model card is the label attached to a model. It spells out what the thing does, what it learned from, and where it shouldn't be used.
- It isn't a brag sheet listing only what the model does well. What it can't do and where it's risky to use is the part that matters most.
- It lists what the model learned from, so you can guess ahead of time which way it's likely to lean.
- Before you use it, check the allowed uses and the terms of use. Being out in the open doesn't mean you can use it however you like.
- The label is self-reported by whoever made it, so some items can simply be missing.
Contents
1The analogy
The front of a snack bag has a big name and a mouthwatering picture. The information you actually need sits on the back. What the ingredients are, how much of what's in there, who should be careful, how to store it — all in small print. To someone with a nut allergy, that one line on the back matters far more than the picture on the front.
Models have a back like this too. The name and the bragging points sit big on the front; what it was made from, who it doesn't suit, and where it's a bad idea to use it sit on the back. Skim only the front and you walk away knowing the name and nothing about the risk. That back-of-the-bag label is a model card.
2In detail
What goes on the label
The first field is what the thing does. What goes in, what comes out, what job it was built for. A model built to summarize text and one built only to sort things into short categories can look alike and still serve completely different jobs.
The second field is recommended and discouraged uses, and it's the most important field on the card. It's common to see a flat warning against using it anywhere a decision lands directly on someone's life — hiring a person, deciding whether to lend money.
The third field is terms of use. Whether commercial use is allowed, whether you can redistribute the output, whether you have to credit the maker. Skip this field, wire the model into a product, and later having to rip it all back out actually happens.
What it learned from is half the story
The card lists what kind of material the training used — where it was gathered, which languages showed up and how much, how recent the material is. Those few lines explain most of a model's personality.
Mostly English material means it can sound awkward in another language; a lot of writing from one region means it answers by that region's default assumptions. Anything after the material was collected, it simply doesn't know about. Read the date on the card and the question "why doesn't it know about something recent" answers itself.
The results of skewed material rarely show on the surface. The more detail a card gives about its material, the more you know ahead of time where to be careful, and you can go test that spot specifically.
A score is a number with strings attached
The card also carries a report card. Which test it was measured against, what score it got. That number is a number from that particular test, not a number from your particular work.
A well-written card also spells out where the score drops. Weak on a certain tone of writing, falls apart on very long text, uneven results for a certain group — that kind of sentence. A card with sentences like that is far more trustworthy than one showing only big, bare numbers.
Safety-side test results sometimes appear too — how it was tuned to decline risky requests, how that was tested. For a model going into something that talks to people, that section is worth reading before the performance numbers.
The order to read a card in
Set the name and the score on the front aside for a moment and start with discouraged uses. If what you're trying to do is on that list, nothing after it matters.
Next, look at the training material and the terms of use, and only then the scores. Then measure it yourself, briefly, on your own data. However well a card is written, it can't stand in for measuring it on your own turf.
A card with a lot of blanks is information too. A card that skips the training material and the limitations is a sign the model itself got less care.
3More precisely
A model card isn't a legally required form — it's a practice that researchers proposed and that spread from there. It gets posted alongside the model file when a model is released, and how much detail goes in each field varies by whoever made it. A related document for the training material side exists too, sometimes called a datasheet.
The label analogy breaks down in a spot. A food label follows a fixed set of rules, and getting the facts wrong carries consequences; a model card is mostly voluntary, so there's a weak mechanism for catching an unfavorable detail left out. A snack's contents also stay exactly what's in the bag, while a model can get quietly updated, so a card might be describing an old version. As safety rules take shape, more places are being asked to keep a document like this on hand. So reading a card is worth doing as much for what isn't written as for what is. If the whole section on training material is missing, or the field for limitations is blank, that absence is telling you exactly how far to trust the model.
4Try it yourself
5Common misconceptions
It's easy to think a model card is a promotional write-up of how well it performs, but actually the more important part is the space listing what it can't do and where not to use it.
It's easy to think a score on the card is the score you'll get on your own work too, but actually the result shifts along with whatever test produced that number.
It's easy to think an openly released model can be used however you like, but actually conditions on use and redistribution are attached more often than not.
7One-line summary
In shortA model card is the label stuck to the back of a model, and reading it well means reading the limits and the terms on the back before the bragging on the front.
Spotted an error or have a better analogy? Suggest an edit · Last updated2026-09-02